fix(policy_engine): skip pipeline-managed guardrails in the response-path guardrail loop

litellm/proxy/utils.py

@@ -2825,6 +2825,7 @@ class ProxyLogging:
         if pipeline_response is not None:
             response = pipeline_response  # rebind-ok: adopt the pipeline's replacement response, same contract as the callback loops below
 
+        pipeline_managed: Final = _pipeline_managed_guardrail_names(data)
         guardrail_callbacks: Final[list[CustomGuardrail]] = []
         other_callbacks: Final[list[CustomLogger]] = []
         try:
@@ -2849,12 +2850,18 @@ class ProxyLogging:
             guardrail_data: Final = _check_and_merge_model_level_guardrails(data=data, llm_router=llm_router)
 
             parallel_guardrails: Final[tuple[CustomGuardrail, ...]] = tuple(
-                callback for callback in guardrail_callbacks if getattr(callback, "run_in_parallel", False)
+                callback
+                for callback in guardrail_callbacks
+                if getattr(callback, "run_in_parallel", False)
+                and not (callback.guardrail_name and callback.guardrail_name in pipeline_managed)
             )
 
             for callback in guardrail_callbacks:
                 # Main - V2 Guardrails implementation
 
+                if callback.guardrail_name and callback.guardrail_name in pipeline_managed:
+                    continue
+
                 if getattr(callback, "run_in_parallel", False):
                     continue
 

tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py

@@ -951,6 +951,32 @@ async def test_post_call_pipeline_pass_runs_once_and_leaves_request_data_untouch
     assert "guardrails" not in data["metadata"]
 
 
+@pytest.mark.asyncio
+async def test_post_call_pipeline_managed_default_on_guardrail_runs_exactly_once(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class CountingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["count"] += 1
+            return None
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [CountingGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=True)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+
+    await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert seen["count"] == 1
+
+
 @pytest.mark.asyncio
 async def test_post_call_pipeline_replacement_response_reaches_caller(
     proxy_logging, make_user_api_key_auth, monkeypatch